bayazidhs04/liltii-v0.2-qwen2.5-0.5b

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Oct 15, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The liltii-v0.2-qwen2.5-0.5b model, developed by bayazidhs04, is a 0.5 billion parameter QLoRA fine-tune of Qwen/Qwen2.5-0.5B-Instruct, featuring a 32,768 token context length. It is specifically designed as an on-device chat companion for the CupCake project, excelling at short, casual, emotion-aware conversations in Bangla, English, and romanized Bangla ("Banglish"). This model's primary differentiator is its optimization for resource-constrained environments like mobile phones, delivering fast, in-character responses with emotion tags for companion applications, and demonstrating superior in-domain quality and speed compared to its base model and other tiny models.

Loading preview...

Overview

liltii-v0.2-qwen2.5-0.5b is a 0.5 billion parameter QLoRA fine-tune of Qwen/Qwen2.5-0.5B-Instruct, developed by bayazidhs04. It serves as the "Liltii" persona for the "CupCake" project, an emotion-aware chat companion designed for on-device use. The model specializes in short, casual, in-character conversations in Bangla, English, and romanized Bangla ("Banglish"), formatting replies with a leading emotion tag and optional actions (e.g., [sleepy] *lies down* ghum...).

Key Capabilities

  • Emotion-aware chat: Generates responses with explicit emotion tags ([happy], [sad], [sleepy], etc.) to drive companion animations.
  • Multilingual support: Handles conversations in Bangla, English, and "Banglish" (romanized Bangla).
  • On-device optimization: Designed for efficient local execution on mid-range phones via llama.cpp GGUF quants, pairing with BLE-connected microcontrollers.
  • Concise responses: Tuned for short, casual turns, averaging 56 characters per reply, ideal for chat companions.
  • Superior in-domain performance: Achieves a BPB of 1.35 on an in-domain Banglish chat corpus, outperforming larger models like Llama-3.2-1B-Instruct (1.52 BPB) and its own base model (1.56 BPB).
  • Faster than base model: Delivers 20.9 tokens/second (Q4_0) compared to 17.1 tokens/second for its base model, due to optimized Q4_0 quantization.

Good For

  • On-device chat companions: Ideal for applications requiring a small, fast, and persona-driven conversational agent on mobile or edge devices.
  • Bangla/Banglish interactive experiences: Suited for engaging users in Bangla, English, or mixed-language chat.
  • Emotion-driven interfaces: Useful for systems where the model's emotional output can control UI elements or animations.
  • Resource-constrained environments: Excellent choice when memory and computational power are limited, such as embedded systems or older smartphones.

Limitations

  • Not suitable for factual QA, coding, math, translation, or high-stakes decision support.
  • Not designed for long-form generation.
  • May reproduce slang or stereotypes from its training data.
  • Lacks RLHF/DPO safety stages; requires host-side output filtering for end-user deployment.

Finetunes
1 models